天风证券：AI存储革命已至，“以存代算”开启存储新纪元

Core Viewpoint - The "storage instead of computing" paradigm emerges as a disruptive technology to overcome computing bottlenecks and storage limitations in AI inference, significantly reducing latency and costs while enhancing throughput [1][2]. Development Background - AI inference has become a key measure of the commercial value of large models, facing challenges such as slow processing and high costs. The "storage instead of computing" technology addresses these issues by migrating vector data from expensive DRAM and HBM to cost-effective SSDs, enabling strategic expansion of storage layers [1]. Core Technology - The "storage instead of computing" Cached Attention technology caches historical KVCache data across HBM, DRAM, and SSD, achieving an 87% reduction in first token latency and a 7.8-fold increase in prefill throughput, leading to a 70% decrease in end-to-end inference costs [2]. Hardware Breakthroughs - Under the "storage instead of computing" paradigm, SSDs evolve from mere data storage to core components in AI inference, requiring high capacity, throughput, and low latency. The AISSD technology will develop in three directions: transitioning to QLC particles, adopting PCIe 5.0/6.0 interfaces with NVMe protocols, and upgrading functionalities towards intelligent solutions [4]. Industry Layout - Major industry players are actively engaging in the core practices of "storage instead of computing," with companies like Huawei and Inspur optimizing storage architectures and cache management for efficient AI inference [5][6]. International Developments - Global giants such as Kioxia, Micron, and Solidigm are pushing for technological iterations and product innovations in AISSD, with QLC+PCIe/NVMe+CXL expected to form the foundation for the next generation of AISSD, transforming SSDs into long-term memory carriers for AI inference [10]. Investment Recommendations - The AI storage revolution is underway, with "storage instead of computing" creating significant opportunities. Companies in storage module manufacturing, storage chips, and distribution/testing are recommended for attention, including Jiangbolong, Demingli, and Zhaoyi Innovation [10].